Network diagnosis framework for bearing faults based on simulation data and subdomain adaptation
In practical industrial environments,there is often a lack of bearing fault data for corresponding operating conditions for model training,which limits the application of deep learning in industrial scenarios.Based on this,two modeling methods were used to generate bearing fault signals,which were used for training the model.Deep sub domain adaptive methods were used to narrow the difference between simulated and real signals,improving the diagnostic accuracy of the model for real signals.Firstly,both mathematical modelling and LS-DYNA-based finite element simulation were used to build a bearing fault simulation model to obtain the simulated acceleration signals of bearing faults that had the same working conditions as the actual scenario.Secondly,to reduce the domain shift between the simulated and real data,the subdomain adaptation method was used to align the global feature distributions and related subdomain feature distributions of the simulated and real data.Finally,the original one-dimensional vibration signal was used as input to implement end-to-end bearing fault classification on the improved residual network(ResNet)model architecture.The bearing fault signals collected by the University of Paderborn were validated as experimental data.The research results indicate that comparing to finite element simulation,the simulated signals generated by mathematical modelling can be more easily transferred to the actual signals,and have a bearing fault identification accuracy of 99.73％in the unlabeled data scenario.This shows that mathematical modelling has greater potential to solve the problem of insufficient fault samples in bearing fault diagnosis,and may be a key technique for building a bridge between real industrial systems and artificial intelligence.
万方数据
维普
